DS202A: Data Science for Social Scientists

Group 8: What Actions Can Be Justified? Are There Regional Differences When It Comes to This?

1. What Actions Can Be Justified?

1.1 Introduction

To answer this research question, we explored and analysed data from the seventh wave of the World Values Survey (WVS). The WVS is an international research endeavour committed to the scientific investigation of values - “social, political, economic, religious, and cultural” - the purpose being to ascertain their temporal stability and impact on societal development (World Values Survey n.d.) (Inglehart et al. 2021). These values are investigated in five year waves, primarily using a “representative comparative social survey” (World Values Survey n.d.). The methodological rigour of the survey design and delivery decreases the probability of suspicion, framing effects, and biases from respondents (Torgler and Schneider 2005; Torgler and Valev 2010). Accordingly, the WVS is globally implemented and recognised in academia and policy, having generated thousands of publications and citations, and been employed by governments and organisations worldwide, including the World Health Organisation and the United Nations (World Values Survey n.d.).

Between 2017 and 2022, 64 surveys were completed on representative, random probability samples of each country’s adult population (Inglehart et al. 2021; Martínez, Sánchez, and Galindo 2022), mostly between 2018-2020 using face-to-face interviews (World Values Survey n.d.). This generated a detailed dataset (Haerpfer et al. 2022) of considerable “geographical and thematic scope”, encompassing 94.5% of the global population and a wide variety of social scientific disciplines (World Values Survey n.d.).

The research question what actions can be justified? is in reference to the actions listed in questions 177 to 195 of the WVS. The actions of interest range from seemingly minor indiscretions, such as avoiding a fare on public transport, to more serious offences like bribery, violence, terrorism, and controversial bioethical issues including abortion, euthanasia, and capital punishment. Acceptability, or justifiability, of such actions will likely vary cross-culturally as the “norms, values, and belief systems” regarding right and wrong differ between cultures (Lee et al. 2004), perhaps due to religious and political ideology, family values, economic status, education level, and other individual, social, economic, and cultural factors. Section 1 of the report will investigate how justifiability of the specified actions can be calculated in answer to the first research question.

RStudio Setup (importing packages and libraries)

library('tidyverse')
library('plotly')
library('tidymodels')
library('moments')
library('car')
library('knitr')
library('kableExtra')
library('dendextend')
library('reshape2')
library('philentropy') # for Jensen-Shannon divergence
library('e1071')       # for skew
library('factoextra')  # for elbow and average silhouette plots
library('dbscan')      # library for DBSCAN
library('ggrepel')
library('parallel') # parallel processing 
library('doParallel')
library('corrplot')

# Maps 
library('maps')
library('countrycode')
library('rnaturalearth')
library('rnaturalearthdata')

# prediction models 
library('rpart.plot')

library('magrittr')

1.2 Loading and Cleaning Data

wvs <- read.csv(unz("WVS.csv.zip", 
                     "WVS.csv"))

Selecting Relevant Variables

# Select relevant demographics columns 
dem_cols <- c("B_COUNTRY", "B_COUNTRY_ALPHA", "N_REGION_ISO", "N_REGION_WVS", "G_TOWNSIZE2", "H_SETTLEMENT", "H_URBRURAL", "O1_LONGITUDE", "O2_LATITUDE")

# Independent variables 
ind_vars <- c("Q1", "Q2", "Q3", "Q4", "Q5", "Q6", "Q260", "X003R2", "Q269", "Q274", "Q275R", "Q287", "Q289")

# Select questions  
wvs_qs <- wvs %>% 
  select(all_of(dem_cols), all_of(ind_vars), Q177:Q195)

# Add country_name column for interpretability 
wvs_qs <- wvs_qs %>%
  mutate(country_name = case_when(
    B_COUNTRY_ALPHA == "NIR" ~ "Northern Ireland", 
    B_COUNTRY_ALPHA == "GBR" ~ "Great Britain", # Resolve GBR/UK/NI disparity 
    TRUE ~ countrycode(B_COUNTRY_ALPHA, 
                      origin = 'iso3c',
                      destination = 'country.name')
  )) %>% 
  relocate(country_name) 

head(wvs_qs) %>% 
  kable()

country_name	B_COUNTRY	B_COUNTRY_ALPHA	N_REGION_ISO	N_REGION_WVS	G_TOWNSIZE2	H_SETTLEMENT	H_URBRURAL	O1_LONGITUDE	O2_LATITUDE	Q1	Q2	Q3	Q4	Q5	Q6	Q260	X003R2	Q269	Q274	Q275R	Q287	Q289	Q177	Q178	Q179	Q180	Q181	Q182	Q183	Q184	Q185	Q186	Q187	Q188	Q189	Q190	Q191	Q192	Q193	Q194	Q195
Andorra	20	AND	20007	20005	3	1	1	1.52	42.51	1	1	1	3	1	4	2	3	2	2	2	3	1	1	1	1	1	1	4	3	2	5	4	1	1	1	1	1	1	1	1	1
Andorra	20	AND	20003	20002	2	2	1	1.53	42.51	1	1	1	4	1	4	1	2	1	0	3	2	1	1	1	1	1	1	9	5	9	9	10	1	9	1	1	1	1	5	1	1
Andorra	20	AND	20003	20002	2	2	1	1.58	42.53	1	2	2	2	3	3	1	2	2	0	3	3	1	1	1	1	1	1	10	6	5	6	6	5	8	1	1	2	1	5	1	2
Andorra	20	AND	20003	20002	2	2	1	1.58	42.53	1	1	1	4	2	4	2	3	2	1	1	4	0	1	1	1	1	1	6	6	4	9	10	3	3	1	1	1	1	7	1	1
Andorra	20	AND	20003	20002	2	2	1	1.58	42.54	1	1	1	3	1	3	1	2	2	0	1	4	0	1	1	1	1	1	8	6	5	7	8	1	5	1	1	1	1	6	1	1
Andorra	20	AND	20006	20006	2	2	1	1.52	42.50	1	3	1	4	1	3	2	3	2	3	1	4	1	10	10	1	1	1	10	1	1	10	10	1	1	1	1	1	1	10	1	1

We select the relevant question numbers for our analysis, Q177:Q195, along with a range of demographic columns to retain the country specific identifiers and geographic information, along with a set of independent variables for prediction in section 2.4. We also retrieve individual country names for ease of interpretation.

Cleaning Data

In the data, N/A and “Don’t Know” responses are coded as -1, -2, -3, and -5. We have decided to replace all instances of these integers with N/A for analysis, as to engage in imputation could introduce potential biases and assumptions that do not accurately reflect respondents. The World Values Survey (WVS) captures a broad range of cultural, societal, and individual perspectives across multiple countries, and these coded responses often reflect uncertainty or lack of knowledge, rather than missing data. Imputing these values could distort the true views of respondents, especially considering the complexity of factors such as national contexts, individual belief systems, and social influences (e.g. the importance of family or religion). Each country may have different interpretations, weights, or legislation regarding certain issues, and so imputing these values could mask regional or demographic differences, necessary for the next part of our analysis – “2. Are There Regional Differences When It Comes to This?”. By rendering these responses as null, we allow for a more accurate representation of the diverse worldviews reflected in the WVS data, ensuring our analysis remains grounded in actual responses provided by participants.

# replace with N/A 
wvs_qs <- wvs_qs %>%
  mutate(across(starts_with("Q"), 
                ~if_else(. >= 1 & . <= 10, ., NA_real_)))

Creating Question Labels

Aids graphical visualisation as there are 19 actions, and using question numbers may lead to confusion.

question_labels <- c(
    "Q177" = "Claiming Unentitled Benefits",
    "Q178" = "Avoiding Transport Fare",
    "Q179" = "Stealing Property",
    "Q180" = "Tax Cheating",
    "Q181" = "Accepting Bribe",
    "Q182" = "Homosexuality",
    "Q183" = "Prostitution",
    "Q184" = "Abortion",
    "Q185" = "Divorce",
    "Q186" = "Sex Before Marriage",
    "Q187" = "Suicide",
    "Q188" = "Euthanasia",
    "Q189" = "Wife Beating",
    "Q190" = "Child Beating",
    "Q191" = "Violence Against Others",
    "Q192" = "Terrorism",
    "Q193" = "Casual Sex",
    "Q194" = "Political Violence",
    "Q195" = "Death Penalty")

Converting the WVS to Long Format (for analysis)

wvs_long <- wvs_qs %>%
  pivot_longer(cols = Q177:Q195, names_to = "question", values_to="response")

# adding question labels
wvs_long <- wvs_long %>%
  mutate(question_label = question_labels[question])

1.3 Exploring Measures of Difference

Descriptive Statistics

To understand justification differences across all actions outlined in the World Values Survey, we begin by exploring descriptive statistics. Since survey responses are recorded on an ordinal scale from 1 (‘Never Justifiable’) to 10 (‘Always Justifiable’), descriptive statistics help summarize the data effectively and facilitate comparisons between questions that elicit similar response patterns.

The Arithmetic Mean: The Central Tendency or Average

• The mean can tell us the average justification score in a country, and for each question

• A higher mean (closer to 10) would suggest that respondents generally justify the action.

• A lower mean (closer to 1) would indicate the action is generally unjustifiable.

• If we compare means, we could show which behaviors are more widely accepted or condemned, and comparison across countries can reveal regional moral differences.

Arithmetic Mean

# compute the mean for each question
mean_data <- wvs_qs %>%
  select(Q177:Q195) %>% 
  summarise(across(everything(), mean, na.rm = TRUE)) 

# convert to long format for plotting
mean_long <- mean_data %>%
  pivot_longer(cols = everything(), names_to = "question", values_to = "mean")  

Why Might We Hesitate to Use the Mean?

• The mean is sensitive to outliers, and if some respondents provide responses at the extreme ends of the scale, or if there are multiple peaks in the distribution, the mean will be skewed.
• The mean cannot not show polarisation, as if 50% of respondents choose 1 (never justified) and the other 50% choose 10 (always justified), the mean would wrongly suggest neutrality.
• The mean does not take the measure of the variable being investigated (the scale is discrete, taking values 1 to 10), and so could return a value of 5.5 or 6.7 

Skewness: A Measure of Asymmetry of a Distribution

In a skewed distribution, the mean is unrepresentative of central tendency.

skew_df <- wvs_qs %>%
  summarise(across(Q177:Q195, ~skewness(., na.rm = TRUE))) %>%
  pivot_longer(everything(), 
               names_to = "question", 
               values_to = "skewness")

• All responses are highly right skewed; for all of the distributions the mode is 1
• skewness: 0.5 < x < 1 is moderately skewed, if x > 1 then highly skewed

This informs us that the mean is an ill-suited measure to the data.

We can an intuitive grouping of actions – some questions have less consensus/are more controversial.

The Median: The Middle Value in a Sorted List (50% of data is below it, 50% is above)

• The median is less sensitive to extreme values (outliers) than the mean.

• A lower median would mean that most people reject the action as unjustifiable.

• A higher median suggests widespread justification of the action.

• Together, if the mean and median differ significantly, we can conclude that the data is skewed.

# calculate the median for each question
median_data <- wvs_qs %>%
  select(Q177:Q195) %>% 
  summarise(across(everything(), median, na.rm = TRUE))

Why Might We Hesitate to Use the Median?

• The median only tells us the middle value, not how spread out the data is.

• If there are two peaks in the distribution, the median can be misleading as it may fall in the middle without reflecting extreme opinions.

The Median + Upper (Q3) and Lower Quartiles (Q1), and the Interquartile Range (IQR)

Investigating the Upper and Lower Quartiles may lead us to more insights into the data.

• These values can detect polarisation

  • A large IQR with a median near the middle suggests polarised views 

  • If Q1 and Q3 are close to the median, most people generally agree on a response.

• These values can enhance the analysis by showing variability and distribution shape

  • If the IQR is small, responses are consistent around the median.

  • If the IQR is large, opinions are more spread out, suggesting greater disagreement.

• Q1 and Q3 can detect skewness, and the direction of skewness 

  • If Q3 is further from the median than Q1, responses are positively skewed – respondents have provided more high scores

  • If Q1 is much further from the median than Q3, responses are negatively skewed – respondents have provided more low scores

Why Might We Hesitate to Use These Values?

• Use of these values can miss important nuances, like the shape or presence of multiple peaks.

• They don’t reflect how dispersed values are in the IQR, ignoring patterns in the distribution.

The Mode: The Most Frequently Occurring Response for a Given Action

• The mode identifies the dominant opinion

• If multiple values appear frequently, the action might be controversial

• If there is a distinct mode value, opinions on the action can be assumed to be homogeneous

The Mode per Question

# function to compute the mode (from ChatGPT)
compute_mode <- function(x) {
  uniq_x <- unique(na.omit(x))  # remove NAs and get unique values
  uniq_x[which.max(tabulate(match(x, uniq_x)))]  # return the most frequent value
}

mode_data <- wvs_qs %>%
  summarise(across(Q177:Q195, compute_mode, .names = "mode_{.col}")) %>%
  pivot_longer(cols = everything(), names_to = "question", values_to = "mode_value") %>%
  mutate(question = stringr::str_remove(question, "mode_")) %>%
  mutate(question_label = dplyr::recode(question, !!!question_labels))

Why Might We Hesitate to Use the Mode?

• The mode ignores the frequency of other responses – only tells us the most common value

• If two or more values occur frequently (multimodal), the mode can be ambiguous.

The mode is homogeneous across all questions and countries, so does not tell us anything meaningful about the data or differences in opinion across questions and countries. We gain a heightened understanding of the data through looking at the second and third most frequent values.

Second and Third Mode Functions

# function to compute the second highest response variable
compute_second_mode <- function(x) {
  x <- na.omit(x)  # remove NAs
  if (length(x) < 2) return(NA)  # return NA if there aren't enough unique values
  
  freq_table <- table(x)  # count occurrences
  sorted_values <- names(sort(freq_table, decreasing = TRUE))  # sort by frequency
  
  if (length(sorted_values) < 2) {
    return(NA)  # return NA if there's only one unique value
  } else {
    return(as.numeric(sorted_values[2]))  # return second most frequent value
  }
}

# function to compute the third highest response variable 
compute_third_mode <- function(x) {
  x <- na.omit(x)  # remove NAs
  if (length(x) < 3) return(NA)  # return NA if there aren't enough unique values
  
  freq_table <- table(x)  # count occurrences
  sorted_values <- names(sort(freq_table, decreasing = TRUE))  # sort by frequency
  
  if (length(sorted_values) < 3) {
    return(NA)  # return NA if there are fewer than three unique values
  } else {
    return(as.numeric(sorted_values[3]))  # return third most frequent value
  }
}

Mode, Second Mode, and Third Mode Values

wvs_mode <- wvs_long %>%
  group_by(question) %>%
  summarise(
    mode = compute_mode(response),
    second_mode = compute_second_mode(response),
    third_mode = compute_third_mode(response)
  )

This graph provides an initial view of the most common responses, but to fully understand respondents’ attitudes, we need to analyse the entire distribution of responses through relative frequency graphs.

Relative Frequency Graphs per Question

relative_freq_data <- wvs_long %>%
  filter(response %in% 1:10) %>%  # ensure responses are within 1-10
  group_by(question, response) %>%
  summarise(count = n(), .groups = "drop") %>%
  group_by(question) %>%
  mutate(percentage = (count / sum(count)) * 100)  # compute percentage

Using the mode alone classifies each action as “Never Justifiable”, but the relative frequency graphs reveal important differences in variation and distribution that the mode fails to capture. For instance, while the mode for Q185: Divorce is 1, only roughly 25% of respondents selected this option, indicating a more diverse range of opinions than the mode suggests. This limitation led us to explore distribution parameters for deeper analysis, enabling pairwise comparisons and comparative judgments as an alternative approach to answering our primary research question: “What Actions Can Be Justified?”

The Variance: How Much Individual Responses Deviate From the Mean from 0 (agreement)

• A low variance means most people gave similar responses, indicating consensus.

• A high variance suggests disagreement – some justify the action while others reject it.

  • This may signal polarisation within a society regarding a particular action.

• It is harder to interpret than standard deviation (not in the unit of the original scores)

• It is influenced by extreme values as outliers can make variance large

• It is not useful alone as variance needs context – knowing how responses are spread is more meaningful than just knowing that they’re spread out.

We will examine the possibility of using the standard deviation as a transformation of the variance.

The Standard Distribution: A Measure of Dispersion in the Same Unit as the Original Scores (1-10)

• This measure is more interpretable than the variance

• A small standard deviation means responses are close to the mean – agreement.

• A large standard deviation means responses are widely spread out – mixed opinions.
  e.g. if a country’s standard deviation for “Q184: Abortion” is low, most people either strongly justify or strongly reject bribery, and if it is high, opinions vary greatly within the population.

# calculate the standard deviation for each question
sd_data <- wvs_qs %>%
  select(Q177:Q195) %>%
  summarise(across(everything(), sd, na.rm = TRUE))

• This value assumes normality, and so is most meaningful when data follows a normal distribution, but moral attitudes among a population often don’t.

• It does not show the direction of skewness towards justification or unjustification, just spread

• It is sensitive to outliers; extreme responses can inflate the standard deviation

The Geometric Mean

In face of the drawbacks of using the arithmetic mean, we explored the method of taking the mean of the log distribution then back-transforming by the exponential (~exp()) to give us the geometric mean of the original distribution. The geometric mean gives more weight to the common lower values. Since the arithmetic mean is pulled up by high values, this should provide a more representative average value.

• It is less sensitive to extreme values than the arithmetic mean because it multiplies values instead of summing them - this is particularly useful if some respondents rated actions as “Always Justifiable” (10) or “Never Justifiable” (1) in an extreme manner, distorting the arithmetic mean.

• We use the log transformation, as it rightly suggests the data is right-skewed

• The geometric mean naturally reduces skewness and gives a more balanced central tendency.

geometric_means <- wvs_qs %>%
  summarise(across(Q177:Q195, ~exp(mean(log(.), na.rm = TRUE)))) %>%
  pivot_longer(cols = Q177:Q195,
              names_to = "question",
              values_to = "geometric_mean")

In comparison to the Arithmetic Mean (the graph is shown under The Arithmetic Mean: The Central Tendency or Average), we see that respondents’ action classifications generally lean towards the lower end of the scale, classifying most actions as unjustifiable. However, Divorce and Sex Before Marriage show higher values, suggesting greater acceptance or more variation in justification across respondents.

mean_comparison <- geometric_means %>%
  left_join(mean_long, by = "question") %>%
  mutate(
    mean_difference = mean - geometric_mean,
    mean_ratio = mean / geometric_mean
  )

There is a large difference between the arithmetic and geometric means, as seen in actions like Sex Before Marriage, Homosexuality, and Divorce. This suggests there are outliers or a skewed distributions. The arithmetic mean is disproportionately affected by these extreme values, making it less reflective of the typical respondent’s view, whereas the geometric mean provides a more balanced and robust measure of central tendency. For questions with lower mean differences (e.g. Terrorism and Stealing Property), the data appears to be more uniform, and while the arithmetic mean may not be as skewed, the geometric mean still provides a more consistent measure of central tendency that is less influenced by any remaining small outliers or variations. It mitigates the influence of extreme responses and better captures the central position of the distribution, particularly when the data is skewed.

# combine the datasets and add normalise mean difference
combined_analysis <- mean_comparison %>%
  left_join(skew_df, by = "question") %>%
  mutate(
    # normalize mean difference to 0-1 scale
    mean_diff_normalized = (mean_difference - min(mean_difference)) / 
                          (max(mean_difference) - min(mean_difference))
  )

head(combined_analysis) %>% kable()

question	geometric_mean	mean	mean_difference	mean_ratio	skewness	mean_diff_normalized
Q177	2.074260	2.973475	0.8992153	1.433512	1.271190	0.4696382
Q178	2.000658	2.841761	0.8411027	1.420413	1.382910	0.4135535
Q179	1.401415	1.818169	0.4167548	1.297381	2.672285	0.0040138
Q180	1.635634	2.233539	0.5979055	1.365550	1.991116	0.1788430
Q181	1.485271	1.966051	0.4807802	1.323699	2.363168	0.0658049
Q182	2.666869	4.051870	1.3850012	1.519336	0.654440	0.9384720

Relationship between Skewness and Normalized Mean Difference

The relationship between normalized mean difference (NMD) and skewness illustrates how the geometric mean offers a direct solution to the challenges posed by skewed distributions. As the geometric mean is less affected by extreme values, it provides a more balanced central tendency measure compared to the arithmetic mean. The observed negative relationship between NMD and skewness can be attributed to the 1-10 scale of the questions, where extreme values contribute to skewness, and the geometric mean mitigates the impact of these extremes, providing a more reliable comparison.

Combining All Descriptive Statistics

# getting second and third mode data to easily merge
second_mode_data <- wvs_qs %>%
  summarise(across(Q177:Q195, compute_second_mode, .names = "mode_{.col}")) %>%
  pivot_longer(cols = everything(), names_to = "question", values_to = "second_mode_value") %>%
  mutate(question = stringr::str_remove(question, "mode_")) %>%  # Clean question names
  mutate(question_label = dplyr::recode(question, !!!question_labels))  # Map to labels

third_mode_data <- wvs_qs %>%
  summarise(across(Q177:Q195, compute_third_mode, .names = "mode_{.col}")) %>%
  pivot_longer(cols = everything(), names_to = "question", values_to = "third_mode_value") %>%
  mutate(question = stringr::str_remove(question, "mode_")) %>%  # clean question names
  mutate(question_label = dplyr::recode(question, !!!question_labels))  # map to labels

# merge statistics for all questions
descriptive_statistics <- mean_long %>%
  left_join(geometric_means, by = "question") %>%
  left_join(median_long, by = "question") %>%
  left_join(mode_data, by = "question") %>%
  left_join(second_mode_data, by = "question") %>%
  left_join(third_mode_data, by = "question") %>%
  left_join(sd_long, by = "question") %>%
  select(-question_label.x, -question_label.y, -question_label)

# view the resulting dataframe
head(descriptive_statistics) %>% kable()

question	mean	geometric_mean	median	mode_value	second_mode_value	third_mode_value	sd
Q177	2.973475	2.074260	1	1	2	5	2.671583
Q178	2.841761	2.000658	1	1	2	5	2.578371
Q179	1.818169	1.401415	1	1	2	3	1.791023
Q180	2.233539	1.635634	1	1	2	3	2.153673
Q181	1.966051	1.485271	1	1	2	3	1.916805
Q182	4.051870	2.666869	3	1	10	5	3.408153

1.4 Which Actions Are Justified?

Q177: Claiming Government Benefits to Which You Are Not Entitled

mean	geometric_mean	median	mode_value	second_mode_value	third_mode_value	sd
2.973475	2.07426	1	1	2	5	2.671583

For Q177, the mean response of 2.97 suggests that most respondents reject the action. The median score of 1 confirms that over half of the respondents place it at the lowest end of the scale. The mode of 1, with secondary and tertiary modes at 2 and 5 respectively, further indicate that the most common stance among respondents is disapproval, though some take a more neutral position. The geometric mean of 2.07, lower than the arithmetic mean, highlights the skewed nature of the responses, reinforcing the general tendency toward unjustifiability. Furthermore, the standard deviation suggests moderate variation, indicating that most respondents reject the action, but a subset expresses more leniency or neutrality. This is consistent in the literature as researchers found a similar mean (2.68), and additionally that income and being male is positively correlated with this action as opposed to age, education, and married status, all of which relate negatively (Navot, Reingewertz, and Cohen 2016). Overall, these statistics place this action firmly within the lower half of the justification scale, with widespread disapproval with some variation in attitudes, as reflected in the relative frequency graph below.

Q178: Avoiding a Fare on Public Transport

mean	geometric_mean	median	mode_value	second_mode_value	third_mode_value	sd
2.841761	2.000658	1	1	2	5	2.578371

For Q178, “Avoiding a Fare on Public Transport”, the mean response of 2.84 suggests that the action is mostly rejected by respondents, a conclusion reinforced by the median of 1, indicating that over half of the responses fall at the lowest point on the scale. The mode of 1, (and secondary and tertiary modes at 2 and 5), further supports the dominant stance of disapproval, though a subset of respondents exhibits neutral or slightly more lenient views. The geometric mean of 2, notably lower than the arithmetic mean, highlights skewness in the distribution, emphasising that the majority lean toward unjustifiability. The standard deviation suggests moderate variation, indicating some divergence in opinion. Overall, while this action is largely deemed unjustifiable, the presence of secondary and tertiary modes, as well as the variation in responses, suggests that attitudes are not entirely uniform, as further illustrated in the relative frequency graph below.

Q179: Stealing Property

mean	geometric_mean	median	mode_value	second_mode_value	third_mode_value	sd
1.81817	1.401415	1	1	2	3	1.791023

For Q179, Stealing Property, the mean of 1.82 and median of 1 indicate overwhelming rejection, with the mode of 1 accounting for over 70% of responses (see the relative frequency graph displayed below), signaling a strong consensus. The geometric mean of 1.40, lower than the arithmetic mean, highlights skewness toward disapproval. Secondary and tertiary modes at 2 and 3 suggest minimal leniency, while the standard deviation of 1.79 reflects low variation. Overall, this action is widely deemed unjustifiable. This conclusion is supported by previous research, in which - out of questions 177, 178, 179, 180, and 181 - stealing property had the lowest mean and was thus viewed as the least justifiable among these actions [Navot, Reingewertz, and Cohen (2016). Further, this un-justifiability extends to stealing in the form of illegal downloading, in which South Asians and North Americans scored the behaviour higher in terms of criminality (Babb 2015).

Q180: Cheating on Taxes If You Have a Chance

mean	geometric_mean	median	mode_value	second_mode_value	third_mode_value	sd
2.233539	1.635634	1	1	2	3	2.153673

For Q180, the mean of 2.23 and median of 1 indicate strong rejection, alongside the mode of 1. The geometric mean of 1.64, lower than the arithmetic mean, shows skewness toward unjustifiability. Secondary and tertiary modes at 2 and 3 suggest some minor variation in responses, while the standard deviation of 2.15 reflects moderate dispersion. Overall, this action is largely deemed unjustifiable, though some respondents show slight divergence in opinion. This divergence may be due to the gender of the respondent as women are more disapproving of tax evasion (Torgler and Valev 2010), and tax cooperation or “tax morale” has increased in Western Europe (Torgler and Schneider 2005).

Q181: Someone Accepting a Bribe in the Course of Their Duties

mean	geometric_mean	median	mode_value	second_mode_value	third_mode_value	sd
1.966051	1.485271	1	1	2	3	1.916805

For Q181, the mean of 1.97 and median of 1 indicate strong disapproval, with the mode of 1 as the most common response. The geometric mean of 1.49, lower than the arithmetic mean, highlights skewness toward unjustifiability. Secondary and tertiary modes at 2 and 3 suggest minor variation, while the standard deviation of 1.92 reflects moderate spread. Overall, this action is widely considered unjustifiable, with limited divergence in responses. Despite the limited divergence observed, researchers have noted gender (Torgler and Valev 2010) and religious differences in the acceptance of bribes, for example the highest justifiability scores were from Muslims, followed by Buddhists, Hindus, Jews, Christians and finally Baha’is (McGee, Benk, and Yüzbaşı 2015).

Q182: Homosexuality

mean	geometric_mean	median	mode_value	second_mode_value	third_mode_value	sd
4.05187	2.666869	3	1	10	5	3.408153

For Q182, the mean of 4.05 suggests a mixed stance, though the median (3) indicates that most responses still lean toward unjustifiability. The mode of 1 remains to be the most common response, but the secondary mode of 10 suggests polarisation, with some respondents deeming the action “Always Justifiable”. The geometric mean of 2.67, significantly lower than the arithmetic mean, highlights skewness toward lower justifiability scores, and a standard deviation of 3.41 reflects high variation, indicating divergent views rather than a clear consensus. Overall, opinions on this action appear divided, as both strong rejection and full justification present. The discrepancies observed in attitudes towards homosexuality are potentially caused by a multitude of socio-cultural factors. Some researchers put negative attitudes down to authoritarianism, inflexible religious beliefs/behaviours, importance of traditional sex roles, less contact with homosexuals (Furnham and Saito 2009), religiosity, and belonging to “the former Soviet and Eastern Bloc, the Muslim World, and sub-Saharan Africa” (Roberts 2019).

Q183: Prostitution

mean	geometric_mean	median	mode_value	second_mode_value	third_mode_value	sd
3.049633	2.11845	1	1	5	2	2.684308

For Q183, the mean of 3.05 suggests respondents generally lean toward unjustifiability, reinforced by the median of 1 and a mode of 1, indicating that the most common response is strong disapproval. However, the secondary mode of 5 suggests some neutrality, however the tertiary mode of 2 further supports a tendency toward rejection. The geometric mean of 2.12, notably lower than the arithmetic mean, highlights skewness toward lower justifiability scores. A standard deviation of 2.68 reflects moderate variation, suggesting that while most respondents consider prostitution unjustifiable, a subset holds more neutral or lenient views. Overall, this action is predominantly seen as unjustifiable, but responses indicate some divergence in opinion. Divergence may result from varying beliefs in marriage (Chon 2015), “political conservatism, authoritarianism and religious fundamentalism”, all of which predict attitudes to prostitution, as well as religiosity, religious denominations and the “survivalist orientation” of one’s country (Stack, Adamczyk, and Cao 2010).

Q184: Abortion

mean	geometric_mean	median	mode_value	second_mode_value	third_mode_value	sd
3.554184	2.430881	2	1	5	10	2.988838

For Q184, the mean of 3.55 suggests a moderate rejection, though the median of 2 indicates most responses are closer to the lower end of the scale. The mode of 1 reflects strong disapproval is most common. However, the secondary and tertiary modes of 5 and 10 respectively indicate significant polarisation, with some respondents showing neutrality and others fully justifying the action. The geometric mean of 2.43, lower than the arithmetic mean, reflects skewness toward lower justification scores. The standard deviation of 2.99 shows high variation, suggesting that views on abortion are divided, with both strong disapproval and justification present in the responses. Overall, opinions on abortion are mixed, with a general tendency toward rejection but notable polarisation in attitudes. This polarisation is well-documented, researchers having noted on the “extreme positions” taken ranging from absolute bans to absolute rights (Tutić Grokša et al. 2025). Abortion is a controversial bioethical issue on which attitudes and even laws differ regionally and internationally. Positive abortion attitudes are more present in women, younger age groups, more educated, and less religious individuals (Loll and Hall 2019). At the cultural level, opposition to abortion is largely founded in religion, but is also related to countries with lower GDP (Tutić Grokša et al. 2025).

Q185: Divorce

mean	geometric_mean	median	mode_value	second_mode_value	third_mode_value	sd
5.007262	3.704647	5	1	5	10	3.202261

For Q185, the mean of 5.01 and median of 5 indicate a neutral stance, but the mode of 1, representing 25% of responses, shows significant disapproval. The secondary mode of 5 suggests greater acceptance, while the tertiary mode of 10 reflects strong justification of divorce. The geometric mean of 3.70, lower than the arithmetic, highlights skewness toward lower justification scores. The high standard deviation of 3.20 suggests diverse opinions on divorce. It is more justifiable than any other action, though opinions remain polarised, with both strong disapproval and justification present, but investigating the relative frequency graph shows most responses, if not 1, “Never Justifiable”, tend to be a value of 5 or above, making it reasonable to conclude that Divorce is a widely justifiable action. Evaluations and rates of divorce are noted to be “systematically linked to cultural values”, specifically the individualism-collectivism dimension, national wealth, and family values (Toth and Kemmelmeier 2009), which explains the variation in justifiability scores.

Q186: Sex Before Marriage

mean	geometric_mean	median	mode_value	second_mode_value	third_mode_value	sd
4.655488	3.206734	5	1	10	5	3.435047

For Q186, the mean of 4.66 and median of 5 suggest a neutral to moderate stance, with the mode of 1 indicating disapproval for almost 35% respondents. The secondary mode of 5 points to neutrality, while the tertiary mode of 10 reflects a strong justification for premarital sex. The standard deviation of 3.44 shows high variation, indicating a range of opinions. Overall, premarital sex is viewed with moderate justification, but responses are polarised, with both absolute disapproval and justification present. Disapproval of premarital sex is rooted in regulation of sexual behaviour (Broude 1975), which differs between families, communities, countries, and cultures which aligns with the variety of opinions observed.

Q187: Suicide

mean	geometric_mean	median	mode_value	second_mode_value	third_mode_value	sd
2.581409	1.8169	1	1	5	2	2.439957

For Q187, the mean of 2.58 and median of 1 suggest strong disapproval of the action, with the mode of 1 representing the most common response is that suicide is “Never Justifiable”, indicating a clear rejection of suicide. However, the secondary mode (5) shows some neutrality. The geometric mean of 1.82, significantly lower than the arithmetic mean, highlights skewness toward strong disapproval. The standard deviation of 2.44 suggests moderate variation, indicating some divergence in opinions. Overall, suicide is largely viewed as unjustifiable, though a minority holds more neutral or lenient views. Some researchers have posited differences in perspectives regarding suicide may lie along the individualism-collectivism dimension (Colucci and Lester 2012).

Q188: Euthanasia

mean	geometric_mean	median	mode_value	second_mode_value	third_mode_value	sd
3.882075	2.627797	3	1	5	10	3.175178

Euthanasia For Q188, the mean of 3.88 and median of 3 suggest moderate rejection, though the mode of 1 indicates a significant portion (~40%) of respondents disapprove. The secondary mode of 5 and tertiary mode of 10 show polarisation, with some respondents taking a neutral stance or strongly justifying euthanasia. The standard deviation of 3.18 suggests high variation, indicating diverse opinions. In comparison to suicide (Q187), where the mean of 2.58 and median of 1 show strong disapproval, euthanasia shows more variation and polarisation. While both actions are viewed largely unjustifiable, euthanasia exhibits greater divergence in opinion, as seen in the higher mean, secondary mode of 5, and tertiary mode of 10, indicating a more divided view than the clear rejection of suicide. This is supported by the fact a previous analysis of the WVS suggested that “approximately half the population approved of euthanasia” (Karumathil and Tripathi 2022). The discrepancies in attitudes towards euthanasia can be traced back to both individual and cultural factors. Those who are older, live in rural regions, and of lower socioeconomic status generally possess more negative views regarding euthanasia (Tutić Grokša et al. 2025). Culturally, Christianity presents more “prohibitive” attitudes, and other religions also influence acceptability e.g., Hinduism and Islam (Karumathil and Tripathi 2022), as well as traditional family values (Lee et al. 1996).

Q189: For a Man to Beat His Wife

mean	geometric_mean	median	mode_value	second_mode_value	third_mode_value	sd
1.858026	1.412235	1	1	2	3	1.867325

For Q189, the mean of 1.86 and median of 1 indicate strong disapproval, with the mode of 1 representing the most common response, reflecting clear rejection of this action. The secondary mode of 2 and tertiary mode of 3 suggest minor variation, with some respondents holding more lenient views. The standard deviation of 1.87 suggests moderate variation, showing that while the majority reject the action, there is some divergence in opinions, but overall, a man beating his wife is widely viewed by respondents as unjustifiable, with strong rejection yet some variation. This variation is observed in research as despite staunch opposition, some countries - “Mali, Serbia, Montenegro, Zambia, Burkina Faso, India, the Philippines, Rwanda, Egypt, and Algeria” - somewhat accept intimate partner violence (Tausch 2019). Disrespect towards traditional family values explained the most variance of acceptability (Tausch 2019).

Q190: Parents Beating Children

mean	geometric_mean	median	mode_value	second_mode_value	third_mode_value	sd
2.831576	1.994555	1	1	5	2	2.538224

For Q190, the mean of 2.83 and median of 1 suggest strong disapproval, with the mode of 1 representing the most common response. However, the secondary mode of 5 and tertiary mode of 2 indicate slight polarisation, with a subset of respondents holding neutral and more lenient views. The standard deviation of 2.54 indicates moderate variation, suggesting diverse opinions on the issue. Compared to the action of a man beating his wife (Q189), where the mean of 1.86 and mode of 1 suggest a stronger consensus against the action, parents beating children shows more variation. While both actions are widely rejected, polarisation occurs in the case of parents beating children, as is reflected in the secondary mode of 5 and a higher standard deviation. This indicates that while the majority of respondents reject the action, there is greater divergence in opinions about whether it can ever be justified, whereas wife beating is overwhelmingly condemned. The increased justifiability observed is likely due to the fact that corporal punishment of children is either rare or a “frequent child-training technique in many societies”, predicted by “dependence on agriculture”, and political/economic complexity (Ember and Ember 2005). The discrepancy observed between the justifiability of child and wife beating is that the former is associated with learning while the latter with punishment (Almosaed 2004).

Q191: Violence Against Other People

mean	geometric_mean	median	mode_value	second_mode_value	third_mode_value	sd
1.96055	1.495263	1	1	2	3	1.858066

For Q191, the mean of 1.96 and median of 1 suggest strong disapproval, with the mode of 1 as the most common response. The secondary mode of 2 and tertiary mode of 3 indicate some leniency, but the majority still firmly reject the action. The standard deviation of 1.86 reflects low variation, but suggests slight divergence in opinions. Overall, this action is largely deemed unjustifiable. While violence against other people shares the strong rejection seen in wife beating (Q189), it shows slightly less variation than both wife beating and child beating (Q190), as indicated by its lower standard deviation. This suggests that while the action is widely condemned, there is less divergence in opinions on violence against others compared to the near-universal rejection of wife beating, and more diverse attitudes towards child beating, a fascinating insight from the data. This is despite the fact violence occurs within “every socio‐economic group, ideology, class, race and ethnic grouping” (Almosaed 2004).

Q192: Terrorism as a Political, Ideological or Religious Means

mean	geometric_mean	median	mode_value	second_mode_value	third_mode_value	sd
1.789475	1.37688	1	1	2	3	1.794037

For Q192, the mean of 1.79 and median of 1 suggest strong disapproval, with the mode of 1 supporting this. The secondary mode of 2 and tertiary mode of 3 indicate some leniency, but the majority still firmly reject the action. The standard deviation of 1.79 shows slight variation, suggesting some divergence in opinions, though terrorism is largely viewed as unjustifiable. The observed variance may result from the fact that terrorism in itself is difficult to define (Martínez, Sánchez, and Galindo 2022), and relative to perspectives (Lee et al. 2004). Individual factors that may result in justification include youth, naivety, poor life satisfaction and health, and religion (Martínez, Sánchez, and Galindo 2022). The rejection of terrorism is consistently expressed, such as by the UN which states such acts have “no justification in our societies” (Martínez, Sánchez, and Galindo 2022), which likely affects scores.

Q193: Having Casual Sex

mean	geometric_mean	median	mode_value	second_mode_value	third_mode_value	sd
3.555538	2.410104	2	1	5	10	3.040943

For Q193, “Having Casual Sex”, the mean of 3.56 and median of 2 suggest moderate rejection, with the mode of 1 indicating strong disapproval at roughly 40% of responses. The secondary mode of 5 and tertiary mode of 10 reveal high polarisation, with some taking a neutral stance or fully justifying the action. The standard deviation of 3.04 indicates high variation, suggesting a diverse range of opinions on casual sex, with some respondents expressing more lenient or neutral views. Overall, while the action is viewed as moderately unjustifiable, there is significant variation in attitudes, rendering the classification of engaging in casual sex as justifiable, somewhat ambiguous.

Q194: Political Violence

mean	geometric_mean	median	mode_value	second_mode_value	third_mode_value	sd
1.987385	1.499409	1	1	2	3	1.920265

For Q194, “Political Violence”, the mean of 1.99 and median of 1 suggest strong disapproval, alongside the mode of 1. The secondary mode of 2 and tertiary mode of 3 indicate slight leniency, but it is still majorly rejected. The standard deviation of 1.92 suggests slight variation, indicating some divergence in opinions, though political violence is largely viewed as unjustifiable. In comparison, to Q192: “Terrorism as a Political, Ideological or Religious Means”, there is a similar pattern, with a low mean of less than 2. The secondary mode of 2 and tertiary mode of 3 also suggest similarly suggest slight leniency, and the standard deviation of 1.79 indicates slight variation. While both actions are strongly disapproved, terrorism tends to show slightly stronger disapproval overall, compared to political violence, which displays a bit more variation in attitudes.

Q195: Death Penalty

mean	geometric_mean	median	mode_value	second_mode_value	third_mode_value	sd
4.098131	2.829311	3	1	5	10	3.182879

For Q195, the mean of 4.10 and median of 3 suggest slight rejection, with the mode of 1, as is the same across all actions, indicating strong disapproval. The secondary mode of 5 and tertiary mode of 10 reveal polarisation, with some respondents being more neutral or fully justifying the action. The standard deviation of 3.18 reflects high variation, indicating diverse opinions on the death penalty, with a general trend toward moderate rejection but substantial divergence in attitudes. The polarity of opinions about capital punishment is a well-documented phenomenon. Those from abolitionist countries are less supportive of it than their retentionist counterparts (McCarthy and Brunton-Smith 2024).

In examining the actions outlined in the World Values Survey, we observe varying levels of justification across different issues. Strong disapproval for actions like claiming unentitled government benefits (Q177) and stealing property (Q179) juxtapose significantly polarised actions, such as homosexuality (Q182), divorce (Q185), and abortion (Q184), likely reflecting cultural and regional differences.

Furthermore, many actions, while semantically aligned, show meaningful differences in societal perception; for instance, Q192 (Terrorism) and Q194 (Political Violence) both relate to violent acts and receive strong rejection, yet political violence demonstrates more variation in opinions compared to terrorism. Similarly, Q189 (wife beating) and Q190 (child beating) show clear condemnation across the dataset, but child beating exhibits more variation than wife beating. The question of euthanasia (Q188) reveals a more divided view, with significant variation in responses. While a substantial portion of respondents express disapproval, a notable minority justifies euthanasia under certain circumstances, leading to greater polarisation in opinions. This contrasts with suicide (Q187), which receives more uniformly negative responses across the dataset.

Overall, certain actions exhibit a clear consensus regarding their unjustifiability, while others reveal substantial variation and polarisation, reflecting the diverse perspectives on controversial issues. These differing viewpoints will be explored further under the section [Are There Regional Differences When It Comes to This?], where we will investigate how and if regional contexts influence the justification of actions across different countries.

1.5 Clustering Actions

In the process of determining the level of justification of 19 actions, clustering provides a structured way to group actions based on their justification patterns across the entire population. By applying clustering methods to our research question – “What Actions Can Be Justified? – clustering helps identify broader patterns in moral reasoning and reveals which actions tend to be justified or rejected together.

Following our initial exploratory data analysis, we suggest that the actions listed could be grouped qualitatively into three broad groups; individual liberties, economics/corruption and violent acts. This is also pointed to by the differing average distributions for each question, as evidenced by the skewness values observed in 1.3. We will explore the relationships between the questions first through comparing Spearman correlations, and then Jensen-Shannon Divergence, to better understand possible groupings of actions based on their response patterns.

1.5.1 Spearman Correlations

Spearman’s correlation provides a suitable method for clustering the questions into meaningful groups. As a non-parametric approach it does not assume a normal distribution, so is suitable for calculating similarity across our skewed data. It is also appropriate for our likert-type 1-10 justifiability scales, handling ordinal data effectively, and is resistant to outliers.

Clustering by Spearman Correlation

cor_matrix <- cor(wvs_qs[, c("Q177", "Q178", "Q179", "Q180", "Q181", 
                           "Q182", "Q183", "Q184", "Q185", "Q186",
                           "Q187", "Q188", "Q189", "Q190", "Q191",
                           "Q192", "Q193", "Q194", "Q195")],
                  method = "spearman",  # Spearman correlation
                    use = "pairwise.complete.obs")

dist_matrix <- as.dist(1 - cor_matrix)

hc <- hclust(dist_matrix, method = "complete")

dend <- as.dendrogram(hc)
dend <- set(dend, "labels", question_labels[labels(dist_matrix)])


plot(hc, 
     hang = -1, 
     labels = question_labels[labels(dist_matrix)])

Outcome at k = 4

1. Claiming Unentitled Benefits, Avoiding Transport Fare, Stealing Property, Tax Cheating, Accepting Bribe, Wife Beating, Violence Against Others, Terrorism, Political Violence
2. Homosexuality, Prostitution, Abortion, Divorce, Sex Before Marriage, Suicide, Euthanasia, Casual Sex
3. Child Beating
4. Death Penalty

1.5.2 Divergence Measures for Clustering

Kullback-Leibler Divergence

Kullback-Leibler divergence is a measure of how much a model probability distribution, Q, is different from a true probability distribution, P. In this case, we could pairwise compare two distributions of justification scores and determine similarity or difference, leaving us with clusters which show actions which seem to converge on how justified, unjustified, or polarised respondents are towards them. Jensen-Shannon divergence builds on this, and is discussed below.

Jensen-Shannon Divergence

Based on the Kullback-Leibler measure, Jensen-Shannon divergence is a symmetric and smoothes version of KL divergence which compares Q and P via an average distribution, M. It has a finite value in the range [0, 1], resulting in it being easier to interpret than KL due to its use of relative entropy. In this case, the similarity of two distributions is greater when the JS distance is closer to zero. Therefore, we have decided to continue with JS divergence, and create a pairwise matrix.

Generate Probability Distributions for Each Question

selected_questions <- wvs_long %>%
  filter(response %in% 1:10) %>%  # keep valid response range (1-10)
  group_by(question, response) %>%
  summarise(count = n(), .groups = "drop") %>%
  group_by(question) %>%
  mutate(probability = count / sum(count))  # convert to probability

Computing Jensen-Shannon Divergence

# unique question pairs
question_list <- unique(wvs_long$question)

# compute pairwise Jensen-Shannon Divergence (JS)
js_results <- expand.grid(Q1 = question_list, Q2 = question_list) %>%
  filter(Q1 != Q2) %>%
  rowwise() %>%
  mutate(JSD = {
    prob_P <- selected_questions %>%
      filter(question == Q1) %>%
      select(response, probability) %>%
      right_join(data.frame(response = 1:10), by = "response") %>%
      replace_na(list(probability = 0)) %>%
      pull(probability)
    
    prob_Q <- selected_questions %>%
      filter(question == Q2) %>%
      select(response, probability) %>%
      right_join(data.frame(response = 1:10), by = "response") %>%
      replace_na(list(probability = 0)) %>%
      pull(probability)
    
    # compute Jensen-Shannon Divergence
    JSD(rbind(prob_P, prob_Q), unit = "log2")
  })

Creating a Matrix

# reshaping js_results into a matrix form
js_matrix <- dcast(js_results, Q1 ~ Q2, value.var = "JSD")

# converting the matrix into a distance object (for clustering)
js_dist <- as.dist(js_matrix[, -1])  # remove the row names (questions themselves)

(Agglomerative) Hierarchical Clustering w/Jensen-Shannon Divergence

Hierarchical Clustering is a machine learning algorithm that groups data into a tree of clusters. Using the agglomerative method, each question is initially its own cluster and the algorithm merges the two most similar clusters (based on the smallest JS distance), and this is repeated until all questions are in a single cluster, and this is displayed in a dendrogram. Our denogram is displayed below.

# perform hierarchical clustering using JSD distance matrix
hc <- hclust(js_dist, method = "complete")

# replace default labels (Q177, Q178, ...) with question labels
hc$labels <- question_labels[hc$labels]

# plot the dendrogram with question labels
plot(hc, main = "Hierarchical Clustering of WVS Questions Based on JSD",
     xlab = "Questions", ylab = "Height", cex = 0.7, 
     labels = hc$labels)

After performing hierarchical clustering, we can cut the dendrogram to define the number of clusters, where each cluster contains questions that are most similar to each other in terms of their response distributions. Here, we determined the optimal number of clusters to be four, aligning with the three broad groups previously suggested, with a single cluster containing “Divorce”, which from the dendrogram is a stark outlier.

# perform hierarchical clustering
hc <- hclust(js_dist, method = "complete")

# cut the dendrogram into clusters
num_clusters <- 4
clusters <- cutree(hc, k = num_clusters)

# ensure the questions have labels
hc_labels <- names(question_labels)[match(hc$labels, names(question_labels))] 
hc$labels <- question_labels[hc_labels] 

# create a data frame of the clustered questions
clustered_questions <- data.frame(Question = hc$labels, Cluster = clusters)

Outcome at k = 4

1. Claiming Unentitled Benefits, Avoiding Transport Fare, Prostitution, Suicide, Child Beating

2. Stealing Property, Tax Cheating, Accepting Bribe, Wife Beating, Violence Against Others, Terrorism, Political Violence

3. Homosexuality, Abortion, Sex Before Marriage, Euthanasia, Casual Sex, Death Penalty

4. Divorce

DBSCAN

DBSCAN (Density-Based Spatial Clustering of Applications with Noise) is a machine learning algorithm that groups data points based on their density (Zhang, Ni, and Xu 2020). Unlike hierarchical clustering, DBSCAN does not require one to specify the number of clusters, k, in advance. Instead, it uses two key parameters: eps (the maximum distance between two points to be considered neighbors) and minPts (the minimum number of points required to form a cluster). The algorithm identifies regions of high density as clusters and labels points in low-density regions as noise or outliers. However, the choice of parameters is critical, as they directly influence the resulting clusters. This is a problem we ran into in our attempt to cluster using DBSCAN, and our method of dealing with this issue is detailed below.

js_matrix_numeric <- js_matrix[,-1]  # remove the first column (question labels)  
diag(js_matrix_numeric) <- 0  # replace diagonal NAs with 0 (pairwise comparison shows no difference)

# finding the point of maximum curvature (to get eps value)
js_matrix_num <- as.matrix(js_matrix_numeric)
distances <- sort(js_matrix_num)  # sort distances
slope <- diff(distances)  # first derivative (slope) - difference between consecutive distances
curvature <- diff(slope)  # second derivative (curvature) - difference between consecutive slopes
max_curvature_index <- which.max(abs(curvature))  # find the index of maximum curvature
max_curvature_value <- distances[max_curvature_index + 1]  # value at that point (because of diff)

dbscan::kNNdistplot(js_matrix_numeric, k = 4) +
  abline(h = max_curvature_value, lty = 2, col = "red") +
  scale_y_continuous(n.breaks = 40)  # Adjust y-axis ticks

NULL

We found the point of maximum curvature to be 0.1775979, however found difficulty when running DBSCAN using this value as it returns one cluster, likely due to the eps value being large.

db <- dbscan::dbscan(js_matrix_numeric, eps = 0.1775979, minPts = 3)
fviz_cluster(db, data = js_matrix_numeric, stand = FALSE,
             ellipse = FALSE, show.clust.cent = FALSE,
             geom = "point", ggtheme = theme_bw(base_size=20))

Finding the Optimal eps Value

As the maximum point of curvature is typically taken as the eps value, but this resulted in one cluster, we used trial and error assigning eps based on the 90th percentile of the curvature values.

# threshold based on the 90th percentile of the curvature values
threshold <- quantile(abs(curvature), 0.90)

# identify the indices of curvature values above the threshold
sharp_increase_indices <- which(abs(curvature) > threshold)

# identify the corresponding distances for these points (index + 1 due to diff)
sharp_increase_values <- distances[sharp_increase_indices + 1]

# plot the kNN distances
library(ggplot2)
p <- dbscan::kNNdistplot(js_matrix_numeric, k = 3) +
  scale_y_continuous(n.breaks = 40)

# add the first abline for the maximum curvature (already identified)
p <- p + abline(h = max_curvature_value, lty = 2, col = "red")

# add additional ablines for other sharp increases
for (i in 1:length(sharp_increase_values)) {
  p <- p + abline(h = sharp_increase_values[i], lty = 2, col = "blue")
}

Determining an eps Value Using minPts = 4 (as it is twice the number of dimensions), and eps = 0.08, we gain three clusters, and two “noise” points. The outcome of which aligns with our intuitive grouping, and with hierarchical clustering.

# Create a vector of the question labels (Q177 to Q195)
question_labels <- c("Claiming Benefits", "Avoiding Transport Fare", "Stealing Property", 
                     "Tax Cheating", "Accepting Bribe", "Homosexuality", "Prostitution", 
                     "Abortion", "Divorce", "Sex Before Marriage", "Suicide", "Euthanasia", 
                     "Wife Beating", "Child Beating", "Violence Against Others", "Terrorism", 
                     "Casual Sex", "Political Violence", "Death Penalty")

# Run DBSCAN
db <- dbscan::dbscan(js_matrix_numeric, eps = 0.08, minPts = 4)

# Plot DBSCAN results
library(factoextra)
fviz_cluster(db, data = js_matrix_numeric, stand = FALSE,
             ellipse = FALSE, show.clust.cent = FALSE,
             geom = "point", ggtheme = theme_bw(base_size = 20))

Outcome at eps = 0.08 and minPts = 4

1. Claiming Unentitled Benefits, Avoiding Transport Fare, Prostitution, Suicide, Child Beating

2. Stealing Property, Tax Cheating, Accepting Bribe, Wife Beating, Violence Against Others, Terrorism, Political Violence

3. Homosexuality, Abortion, Euthanasia, Casual Sex, Death Penalty

Noise: Divorce, Sex Before Marriage

Discussion of the Final Clusters

The Spearman correlation, hierarchical clustering, and DBSCAN provide valuable insights into how different actions are grouped based on their justification distributions. Using Jensen-Shannon divergence, hierarchical clustering with k = 4 revealed four distinct clusters, with Divorce standing out as an outlier, reinforcing its unique position in moral attitudes compared to other actions. However, Spearman correlation identifies Child Beating and the Death Penalty as ‘outliers’ (as they are in their own distinct cluster). Meanwhile, DBSCAN, which identifies clusters based on density rather than predefined groupings, produced three main clusters with two noise points, one of which being Divorce, aligning with our intuition and hierarchical results. The consistency between hierarchical clustering and DBSCAN suggests that the identified groupings reflect meaningful patterns in justification attitudes rather than being an outcome due to a particular clustering approach. The presence of noise points in DBSCAN highlights actions with highly variable or ambiguous justifiability, which may indicate deep societal divides or cultural variation. These findings set the stage for further analysis into regional variations and socio-cultural influences on moral attitudes.

2. Are There Regional Differences When It Comes to This?

This section aims to investigate regional differences in values, based on the justifiability across countries of the set of actions discussed in Section 1, in order to answer the question.

In answer to the question “Can we set universally acceptable criteria for what is right, just, good, or evil for all human behaviour and judgment?” researchers stipulate that beliefs, standards, norms, and values differ cross-culturally (Lee et al. 2004), therefore the justifiability of the aforementioned “dishonest or illegal activities” (Torgler and Valev 2010) should vary regionally, between different countries and cultures. For example, the “individualism-collectivism dimension” has been applied in explaining and predicting attitudinal and behavioural differences in which individualistic countries - like Australia, USA, and Western Europe (Green, Deschamps, and Páez 2005) - sanction the pursuit of self-interest, suggesting leniency regarding these actions, whereas “collectivist societies” adhere more to traditional social conventions, therefore would likely disapprove (Toth and Kemmelmeier 2009). However, countries and cultures differ on a multitude of dimensions, such as religiosity and religious beliefs (Roberts 2019), political liberalism and conservatism (McGee, Benk, and Yüzbaşı 2015), national wealth (Stack, Adamczyk, and Cao 2010), and family values (Lee et al. 1996) (Toth and Kemmelmeier 2009), therefore the potential regional differences in attitudes to Q177-195 are arguably more complex than simple individualism-collectivism classifications.

Methodology

To conduct comparative analysis at the country level requires a metric to represent the aggregate responses from within countries. As discussed previously, we recognise the limitations of these averages, especially in the presence of such highly skewed data. As such, in Section 2.1 we first work with the geometric mean of the distribution of responses for each question per country, which provides a partial remedy to the misrepresentation of the arithmetic mean. This allows for clear initial visualisation of national variation, which we plot, for two questions, onto a world map to better understand how the responses vary by geographic region.

Following this, in Section 2.2, we calculate the Jensen-Shannon Divergence (JSD) for each country-question pair, which allows for comparisons using the entire distributions of responses per country. JSD pairs are used to establish (dis)similarities between countries, then fed into a hierarchical clustering to group countries by response patterns. We fit these clusters to a world map and discuss how our observations regarding regional difference fit with the regional values literature.

Finally, we provide some preliminary explanations for potential regional differences in justifiability, by testing the significance of several predictor variables also drawn from the WVS survey.

2.1 Country Means

The clearest way to observe inter-national variation is comparison of national average figures. Therefore, we begin with an overview of mean national responses to the justifiability of each action.

2.1.1 Geometric Mean per Question, per Country

geometric_means <- wvs_qs %>%
  group_by(country_name) %>%
  mutate(B_COUNTRY_ALPHA = first(B_COUNTRY_ALPHA)) %>%
  summarise(
    B_COUNTRY_ALPHA = first(B_COUNTRY_ALPHA), 
    across(Q177:Q195, ~exp(mean(log(.), na.rm = TRUE)))
    )
  # Create long format df 
geometric_means_long <- geometric_means %>% 
  pivot_longer(cols = -c(country_name, B_COUNTRY_ALPHA),
              names_to = "Question",
              values_to = "GEOMETRIC_MEAN")

This plot captures the mean justifiability of each action, with countries plotted along the x-axis. Responses show significant divergence between countries, particularly for certain questions, helping to identify moral consensus and divergence across countries for each action. Actions such as “Violence Against Others” show low justification scores across most countries, indicating broad societal disapproval. Conversely, actions like “Divorce” exhibit higher justification scores in some countries but remain controversial in others (e.g. Nigeria, Kyrgyzstan). Whilst the clearest patterns are the central horizontal grouping of more ‘controversial’ questions discussed in Section 1, the vertical patterns show, for instance, that Pakistan has very low scores (1 - 2) for all questions, whereas Germany has mean justifiability of up to 8 for some questions. Germany is more supportive of various actions, having experienced an increase in support for euthanasia (Tutić Grokša et al. 2025). Similarly, homosexuality has experienced a statistically significant “upswing” in its acceptance in Western countries however not in the “Muslim World” (Roberts 2019) which includes Pakistan. This national variation provides motivation for regional comparisons especially with respect to the set of high-variance questions, which appear to generate significant variation across nationalities.

Some cells display NaN (Not a Number) values, indicating missing data for certain countries on specific questions. These missing values could arise due to survey non-response captured by our data cleaning, or the exclusion of certain questions in some countries (Wife Beating is not asked in the UK, for instance). As mentioned, we will not impute data. To ensure meaningful comparisons we will note these values, keeping them in the data when they do not impact certain analyses, but noting this in our results.

2.1.2 National-Global Mean Divergence

This plot utilises the same data from the previous, but highlights more starkly the relative differences between countries. Following the vertical patterns emphasises positive and negative deviations concentrated to certain countries. Again, Pakistan is has -2 or -3 deviation from global average for several questions, whilst Germany is generally positively deviated from the global average.

2.1.3 Mapping Means

A helpful way to explore the potential regional patterns in this national variation is to examine differences geographically. For this, we use the maps and countrycode packages to match the B_COUNTRY_CODE variable to internationally recognised iso3c codes and retrieve national latitude and longitudinal values. Using the world map from maps we plot the national mean scores for the justifiability of homosexuality and the death penalty; two questions with significant national variation in responses.

Homosexuality

# Exclude territories without values in world map data, that are not present in the WVS survey 
territories_to_exclude <- c("Barbuda", "Canary Islands", "Micronesia", 
                            "Heard Island", "Chagos Archipelago", "Siachen Glacier",
                            "Saint Martin", 
                            "Bonaire", "Sint Eustatius","Saba", "Madeira Islands", 
                            "Azores", "Ascension Island", "Grenadines", 
                            "Virgin Islands", "Kosovo")


# Country codes
world_map <- map_data("world") %>%
  filter(!region %in% territories_to_exclude) %>%
  mutate(iso3c = countrycode(region, 
                            origin = 'country.name',
                            destination = 'iso3c'))

# Joining justification values with the map data
map_data <- left_join(world_map, geometric_means, 
                      by = c("iso3c" = "B_COUNTRY_ALPHA"))

# Map 
ggplot(map_data, aes(x = long, y = lat, group = group, fill = Q182)) +
  geom_polygon(color = "white", size = 0.1) +
  #scale_fill_viridis_c() +  # or another color scale
  scale_fill_gradientn(colors = c("#ff4242", "#ffd449", "#4caf50")) +
  theme_minimal() +
  theme(legend.position = "bottom") +
  labs(fill = "",
       title = "Where Is Homosexuality Considered Justified?")

Homosexuality, whilst broadly justified across the Western World or Global North countries; the USA, Canada, the UK, northern Europe, Australia and New Zealand, is considered less justifiable across South American countries, very unjustifiable across Russia and Asia, and is often never justified on average in the African countries surveyed. The acceptance of homosexuality has risen across many Western countries (Roberts 2019)

Death Penalty

Retentionist nations (e.g., China) are more accepting and supportive of capital punishment compare to their retentionist counterparts (e.g., Germany) (McCarthy and Brunton-Smith 2024). However, the degree of support does vary within the retentionist-abolitionist groups, for example, the UK is an abolitionist nature which shows greater support for the death penalty, whereas Indonesia, a retentionist nation, is more “critical” of the practice (McCarthy and Brunton-Smith 2024).

2.2 National Divergence

Since grouping countries based on a single mean value hides a large amount of variation in the underlying data, and might wrongly characterise certain countries attitudes, we turn to the JS Divergence (explained in more detail in Section 1.6.2) to measure the similarity between the national distributions.

Based on our findings in Section 1, we are interested in the regional variation in responses to a set of questions, so will take the average JS divergence across these questions per country. For this analysis, we focus on a set of (broadly defined) issues of individual autonomy or ‘post-materialist values’ (Inglehart, 2009) represented within the questions, as opposed to those focussed on economic or violent actions, these are:

Homosexuality

Abortion

Divorce

Sex before marriage

Suicide

Euthanasia

Casual sex

These are also the questions with the clearest national variation and least uniform distributions, which show the most meaningful variation across countries; we also expect responses to them to move broadly in the same direction, as they appear to act together (albeit not perfectly) as an indicator of progressive liberal individualism and secular attitudes.

Though it would be interesting to compare country distributions across all questions individually, we do not have the space in this paper to do so. If we included the whole set of questions then distributions moving in different directions, representing attitudes to different dimensions, would make comparisons less meaningful.

2.2.1 Calculate JS Divergence

Prepare Data for Calculation

# Select questions for similarity calculation 

lib_qs <-  c(
    "Q182", # = "Homosexuality"
    "Q184", # = "Abortion"
    "Q185", # = "Divorce"
    "Q186", # = "Sex Before Marriage"
    "Q187", # = "Suicide"
    "Q188", # = "Euthanasia"
    "Q193"  # = "Casual Sex"
)
    
#all_qs <- c("Q177","Q178","Q179","Q180","Q181","Q182","Q183","Q184","Q185","Q186",
#            "Q187","Q188","Q189","Q190","Q191","Q192","Q193","Q194","Q195")

# df for JS
wvs_select <- wvs_qs %>% 
  select(country_name, B_COUNTRY_ALPHA, lib_qs)

# Long df for JS 
wvs_select_long <- wvs_select %>% 
  pivot_longer(
    cols = Q182:Q193, 
    names_to = "question", 
    values_to = "response"
    )

Calculate probability distributions by country and question

ctry_dstr <- wvs_select_long %>%
  filter(response %in% 1:10) %>%  
  group_by(country_name, B_COUNTRY_ALPHA, question, response) %>%
  summarise(count = n(), .groups = "drop") %>%
  group_by(country_name, B_COUNTRY_ALPHA, question) %>%
  mutate(probability = count / sum(count))

JS Divergence

We perform the computation using parallel processing with detectCores function in the parallel package, and registerDoParallel in the doParallel package for computational efficiency.

# Parallel processing
n_cores <- detectCores() - 1  # Leave one core free
cl <- makeCluster(n_cores)
registerDoParallel(cl)

# List of countries
country_list <- unique(ctry_dstr[c("country_name", "B_COUNTRY_ALPHA")])

# Pairs of countries
country_pairs <- expand.grid(
  Country1 = country_list$country_name, 
  Country2 = country_list$country_name,
  stringsAsFactors = FALSE
  ) %>%
  filter(Country1 < Country2) %>% 
  # country codes
    left_join(country_list, by = c("Country1" = "country_name")) %>%
  rename(B_COUNTRY_ALPHA1 = B_COUNTRY_ALPHA) %>%
  left_join(country_list, by = c("Country2" = "country_name")) %>%
  rename(B_COUNTRY_ALPHA2 = B_COUNTRY_ALPHA)

# Computation of JS divergence
country_js_results <- foreach(i = 1:nrow(country_pairs), 
                            .combine = rbind,
                            .packages = c('dplyr', 'philentropy')) %dopar% {
  
  Country1 <- country_pairs$Country1[i]
  Country2 <- country_pairs$Country2[i]
  B_COUNTRY_ALPHA1 <- country_pairs$B_COUNTRY_ALPHA1[i]
  B_COUNTRY_ALPHA2 <- country_pairs$B_COUNTRY_ALPHA2[i]
  
# Probability distributions for Country1
  prob_P <- ctry_dstr %>%
    filter(country_name == Country1)
  
# Probability distributions for Country2
  prob_Q <- ctry_dstr %>%
    filter(country_name == Country2)
  
# Average JS divergence across all questions
  question_list <- unique(prob_P$question)
  jsd_value <- mean(sapply(question_list, function(q) {
    p1 <- prob_P %>% 
      filter(question == q) %>% 
      arrange(response) %>% 
      pull(probability)
    
    p2 <- prob_Q %>% 
      filter(question == q) %>% 
      arrange(response) %>% 
      pull(probability)
    
    if(length(p1) > 0 && length(p2) > 0) {
      JSD(rbind(p1, p2), unit = "log2")
    } else {
      NA
    }
  }), na.rm = TRUE)
  
  # df results
  data.frame(
    Country1 = Country1,
    Country2 = Country2,
    B_COUNTRY_ALPHA1 = B_COUNTRY_ALPHA1,
    B_COUNTRY_ALPHA2 = B_COUNTRY_ALPHA2,
    JSD = jsd_value
  )
}

# Stop the pp cluster
stopCluster(cl)

# Results to tibble 
country_js_results <- as_tibble(country_js_results)

head(country_js_results) %>% kable()

Country1	Country2	B_COUNTRY_ALPHA1	B_COUNTRY_ALPHA2	JSD
Andorra	Argentina	AND	ARG	0.0836321
Andorra	Armenia	AND	ARM	0.3730577
Argentina	Armenia	ARG	ARM	0.2213274
Andorra	Australia	AND	AUS	0.0265729
Argentina	Australia	ARG	AUS	0.0859024
Armenia	Australia	ARM	AUS	0.3863945

2.2.2 Similarity

Comparing the similarity of countries’ responses to these questions provides a preliminary understanding of how countries might be grouped based on values, and how this might align with regional difference.

Create Matrix

Using the JS divergence values, we create an empty matrix using matrix() and loop each row of country_js_results to fill in the matrix. A previous, simpler method using as.matrix() led to assymetry in the matrix.

# Convert to (distance) matrix for visualisation
countries_for_matrix <- sort(unique(c(country_js_results$Country1, country_js_results$Country2)))


js_matrix <- matrix(0, 
                   nrow = length(countries_for_matrix), 
                   ncol = length(countries_for_matrix),
                   dimnames = list(countries_for_matrix, countries_for_matrix))

for(i in 1:nrow(country_js_results)) {
  js_matrix[country_js_results$Country1[i], 
            country_js_results$Country2[i]] <- country_js_results$JSD[i]
  js_matrix[country_js_results$Country2[i], country_js_results$Country1[i]] <- 
    country_js_results$JSD[i]
}

Most Different Countries

data.frame(
  row = row(js_matrix)[lower.tri(js_matrix)],
  col = col(js_matrix)[lower.tri(js_matrix)],
  value = js_matrix[lower.tri(js_matrix)]
) %>%
  filter(row != col) %>%  # Remove self-pairs
  arrange(desc(value)) %>%
  distinct(value, .keep_all = TRUE) %>%  # Remove duplicates
  head(5) %>%  # Get top 5
  mutate(
    Country_1 = rownames(js_matrix)[row],
    Country_2 = colnames(js_matrix)[col],
    JS_Divergence = round(value, 4)
  ) %>%
  select(Country_1, Country_2, JS_Divergence) %>%
  kable(caption = "Top 5 Most Divergent Country Pairs",
        align = c('l', 'l', 'c'))

Top 5 Most Divergent Country Pairs

Country_1	Country_2	JS_Divergence
South Korea	Libya	0.7777
Netherlands	Libya	0.7507
Libya	Canada	0.7296
Libya	Czechia	0.7277
Vietnam	Libya	0.7218

South Korea and Libya have the greatest divergence in average distribution across the justifiability of this set of actions. But all of the 5 most divergent pairs include Libya, suggesting it has significantly unique responses compared to other countries.

Most Similar Countries

data.frame(
  row = row(js_matrix)[lower.tri(js_matrix)],
  col = col(js_matrix)[lower.tri(js_matrix)],
  value = js_matrix[lower.tri(js_matrix)]
) %>%
  filter(row != col) %>%  # Remove self-pairs
  arrange(value) %>%      # Changed to arrange by ascending order
  distinct(value, .keep_all = TRUE) %>%  # Remove duplicates
  head(5) %>%  # Get top 5
  mutate(
    Country_1 = rownames(js_matrix)[row],
    Country_2 = colnames(js_matrix)[col],
    JS_Divergence = round(value, 4)
  ) %>%
  select(Country_1, Country_2, JS_Divergence) %>%
  kable(caption = "Top 5 Most Similar Country Pairs",  # Updated caption
        align = c('l', 'l', 'c'))

Top 5 Most Similar Country Pairs

Country_1	Country_2	JS_Divergence
Peru	Bolivia	0.0058
Mexico	Guatemala	0.0077
United States	Northern Ireland	0.0081
Pakistan	Indonesia	0.0088
Ecuador	Bolivia	0.0092

The most similar countries are Peru and Bolivia, which are close neighbours in the same region. Similarly, Mexico and Guatemala share a border and are considered closely aligned, with similar religious makeup. Religious beliefs and religiosity are significant determinants of attitudes and values. The fact that the majority of Latin America identify as Chrisian, 69% of adults as Catholic (Wormald 2014), likely contributes to the observed similarities between these countries.

2.2.3 MDS Plot

Below is a multidimensional scaling plot, which provides a visualisation of the differences between countries within a 2 dimensional space, using the JS divergence measure, such that countries closer together are the more similar. The two dimensions do not have meaning in themselves, but are constructed such that the relative distances between countries are preserved, allowing us to view the space that will be analysed in the following section by the hierarchical clustering function.

#JS matrix to distance object
js_dist <- as.dist(js_matrix)

# MDS
mds_result <- cmdscale(js_dist, k=2)
mds_df <- as.data.frame(mds_result)
colnames(mds_df) <- c("Dim1", "Dim2")
mds_df$country <- rownames(js_matrix)

# Plot
ggplot(mds_df, aes(x=Dim1, y=Dim2, label=country)) +
  geom_point() +
  # text options for legibility 
  geom_text_repel(
    size = 2.5, 
    max.overlaps = 100, 
    force = 10, 
    force_pull = 0.1, 
    point.padding = 0.1) +
  theme_minimal() +
  labs(title="MDS Plot of Countries based on JS Divergence")

We can observe several interesting patterns from this initial plot; Libya, the Netherlands, South Korea, and to a lesser extent Ethiopia and Myanmar (Burma) are outliers from the central cluster. The Netherlands may be considered an outlier as it is arguably the most progressive of the countries, described as a “culture of tolerance” in which many of the measured actions are permissible, including abortion (Loll and Hall 2019), euthanasia (Karumathil and Tripathi 2022), and, a previously discussed action, prostitution (Stack, Adamczyk, and Cao 2010). Conversely, sub-Saharan African countries, such as Ethiopia, are associated with declining rates of “positive attitudinal change” (Roberts 2019).

Western, Global North countries appear to be grouped in the upper left of the space, with Eastern European and Latin American countries closer to the centre, and a set of Middle Eastern, and Asian countries on the right. The central cluster of some Eastern European and Latin American countries may cluster due to heightened religiosity as the former “experienced a religious resurgence” post-Soviet Communism (Inglehart et al. 2021) and the majority of Latin America is religious, specifically Catholic (Wormald 2014). The Global North or Western countries are considered more culturally individualistic (Green, Deschamps, and Páez 2005) whereas, Middle Eastern and Asian countries share more collectivist traits (Karumathil and Tripathi 2022), and religion in some cases, which may explain the respective similarities of their attitudes.

2.3 Clustering Countries

To interrogate the similarity across countries and their possible regional divergence, we perform hierarchical clustering to group the countries. Having tried DBSCAN, we found hierarchical clusters to have more meaningful results and be more interpretable through the visualisation of the splits in a dendrogram diagram. This method also allowed us to alter the single k parameter easily, to compare the effect of different numbers of clusters.

2.3.1 Perform Hierarchical Clustering

We first use k = 7 clusters, following some consensus in the literature on a world split into 7 cultural regions. In an analysis of WVS data these “world regions” were specified as “the West (including Western Europe and European settler colonies in North America and the Pacific), Latin America and the Caribbean, the former Soviet and Eastern Bloc, the Muslim World, sub-Saharan Africa, South and Southeast Asia, and East Asia” (Roberts 2019).

We avoid using "complete" clustering, as it is sensitive to outliers, which we have already observed in our data. "average" produces a large number of single country clusters and groups various others together, which is not meaningful, so we use ward.D2, which minimises within-cluster variance and creates more balanced and evenly-sized clusters.

# hierarchical clustering
hc <- hclust(as.dist(js_matrix), method = "ward.D2") # (using distance matrix)

# Number of clusters
k <- 7

# Cut tree 
clusters <- cutree(hc, k)

# Data frame of cluster assignments, with B_COUNTRY_ALPHA attached 
cluster_df_7 <- data.frame(
  Country = names(clusters),
  Cluster = clusters
) %>%
  distinct() %>% 
   left_join(
    country_js_results %>%
      select(Country1, B_COUNTRY_ALPHA1) %>%
      distinct() %>%
      rename(Country = Country1, 
             B_COUNTRY_ALPHA = B_COUNTRY_ALPHA1),
    by = "Country"
  ) %>%
  arrange(Cluster, Country)

# Print 
cluster_df_7 %>%
  group_by(Cluster) %>%
  summarise(Countries = paste(Country, collapse = ", ")) %>%
  kable(col.names = c("Cluster", "Countries"),
        caption = "Countries Grouped by Cluster") %>%
  kable_styling(bootstrap_options = c("striped", "hover"),
                full_width = FALSE)

Countries Grouped by Cluster

Cluster	Countries
1	Andorra, Australia, Canada, Czechia, Germany, Great Britain, Netherlands, New Zealand
2	Argentina, Chile, Greece, Hong Kong SAR China, Japan, Macao SAR China, Mongolia, Northern Ireland, Russia, Serbia, Slovakia, South Korea, Taiwan, Ukraine, United States, Uruguay, Vietnam
3	Armenia, Bangladesh, China, India, Indonesia, Iran, Iraq, Jordan, Kyrgyzstan, Lebanon, Maldives, Myanmar (Burma), Nigeria, Pakistan, Romania, Tunisia, Turkey, Uzbekistan, Zimbabwe
4	Bolivia, Brazil, Colombia, Cyprus, Ecuador, Guatemala, Kazakhstan, Kenya, Malaysia, Mexico, Morocco, Nicaragua, Peru, Philippines, Puerto Rico, Singapore, Thailand
5	Egypt, Ethiopia, Libya
6	Tajikistan
7	Venezuela

With 7 clusters, there emerge (1) a set of European and Western countries, (2) A range of Latin American, East Asian, and Eastern European countries, along with Northern Ireland and the United States, (3) Middle Eastern, African, and South Asian countries, (4) A range of South American, African, and some East Asian countries, then (5) North African Libya and Egypt, and Ethiopia, and (6) and (7) as apparent outliers.

Whilst these groupings appear to follow some aforementioned expected cultural regional attitudes, they could be more meaningful, and we see groupings of countries with very different histories, religions and values, so we plot a dendrogram to see the structure of the clusters produced.

Cluster Dendrogram, K = 7

Visualising the clustering process is helpful to interpret the dissemination of each group of countries. The dendrogram using base R plot() shows the levels at which the clusters are split. We add boxes around clusters with rect.hclust(). higher values of k to further differentiate across the clusters.

plot(hc, hang = -1, cex = 0.6)

rect.hclust(hc, k, border="red")

A significant split at a lower level of the dendrogram splits off South Korea, Vietnam, Mongolia, Taiwan, Hing Kong and Macao from a group of mostly Eastern European countries. This suggests a close regional association that is not captured by the k = 7 parameter, so we increase the value of k to capture these more nuanced regional differences in responses. The cluster dendrogram allows us to know that these clusters already exist in the responses to the questions, and provides justification for altering our method to account for these.

K = 10

Countries Grouped by Cluster

Cluster	Countries
1	Andorra, Australia, Canada, Czechia, Germany, Great Britain, Netherlands, New Zealand
2	Argentina, Chile, Greece, Japan, Northern Ireland, Russia, Serbia, Slovakia, Ukraine, United States, Uruguay
3	Armenia, China, India, Indonesia, Iran, Iraq, Jordan, Kyrgyzstan, Lebanon, Maldives, Nigeria, Pakistan, Romania, Tunisia, Turkey, Uzbekistan, Zimbabwe
4	Bangladesh
5	Bolivia, Brazil, Colombia, Cyprus, Ecuador, Guatemala, Kazakhstan, Kenya, Malaysia, Mexico, Morocco, Nicaragua, Peru, Philippines, Puerto Rico, Singapore, Thailand
6	Egypt, Ethiopia, Libya
7	Hong Kong SAR China, Macao SAR China, Mongolia, South Korea, Taiwan, Vietnam
8	Myanmar (Burma)
9	Tajikistan
10	Venezuela

2.3.2 Regions

Our clustering suggests significant geographic regional clustering of countries on their responses to the justifiabilty of these actions, but manually grouping the questions is not easily interpretable. To compare these clusters to existing geographic regions, we plot them as 10 colours on a world map (using ne_countries for a more detailed map), again linking with iso_a3 codes.

# world map 
world <- ne_countries(scale = "medium", returnclass = "sf")

# 
world_clusters <- world %>%
  left_join(cluster_df_10, by = c("iso_a3" = "B_COUNTRY_ALPHA"))

# map
ggplot(data = world_clusters) +
  geom_sf(aes(fill = factor(Cluster))) +
  scale_fill_brewer(palette = "Set3", na.value = "grey95", name = "Cluster") +
  theme_minimal() +
  theme(
    panel.grid = element_blank(),
    plot.background = element_rect(fill = "white"),
    legend.position = "bottom"
  ) +
  labs(title = "Country Clusters Based on Value Justifiability")

Mapping the country clusters reveals distinct patterns in attitudes towards personal autonomy and moral behaviors (homosexuality, abortion, divorce, premarital sex, suicide, euthanasia, and casual sex) across countries. While geographic proximity appears to be a strong determinant of these moral attitudes, with clear regional patterns emerging, the clusters do not align perfectly with continental groupings. The analysis identifies ten distinct clusters: Western liberal democracies form a cohesive group (Cluster 1), while a mixed cluster of developed and middle-income nations (Cluster 2) bridges traditional Western and non-Western attitudes. A large Islamic-influenced cluster (Cluster 3) spans from North Africa through the Middle East to South Asia, while East Asian nations form their own distinctive group (Cluster 7). Latin American, Southeast Asian, and African nations largely group together (Cluster 5), suggesting similar moral frameworks despite geographic distance, and insinuating there to be a shared underlying explanation. Several single-country clusters (Bangladesh, Myanmar, and Tajikistan) indicate unique moral attitude profiles that don’t align with their neighbors. Three North/East African nations form a distinct group (Cluster 6), while the remaining countries distribute across the clusters in ways that suggest the influence of historical, religious, and developmental factors beyond simple geographic proximity.

These results suggest a significant regional variation in attitudes toward the justifiability of the specified actions, however, further analysis could employ a more complete set of national data to more meaningfully understand global regional differences. There is a particular lack of data across Africa and Europe, which could allow for an interesting lower-level regional analysis on Europe, which generally has the highest variation, or “plurality” (Karumathil and Tripathi 2022), in cultural attitudes.

2.4 Prediction

2.4.1 Overview and Rationale for Prediction

Having illustrated that the justifiability of the specified actions varies between countries and thus regionally, we investigated the underlying mechanisms of this difference by determining the most important predictors of divorce. In order for sample sizes to be large enough to generate meaningful predictions, we did not compare at the national level.

The rationale for divorce as the outcome variable first considers the fact divorce is only illegal in two countries (the Philippines and Vatican City), which would influence justifiability as explicit social norms (e.g., laws) impact beliefs and values. Further, evaluations of divorce are multifaceted, impacted by cultural individualism, societal wealth, family and religious values (Toth and Kemmelmeier 2009), therefore are predicted by a multitude of individual characteristics. Finally, divorced individuals (who supposedly justify this action) are more supportive to other specified actions, such as abortion (Loll and Hall 2019), suggesting predictors may have some degree of generalisability to other actions of interest.

Classification is defined as a “predictive modelling process” which categorises data into specified classes (Belcic 2024), in this instance a binary factor variable representing unjustifiable (0) justifiable (1) actions, specifically divorce. This decision was based on the choice of similar research to recode these WVS variables scored 1-10 to a 0 and 1 “dichotomy” in which 0 represented scores of 1 - actions deemed never justifiable - and 1, which comprised all other scores in which actions were, to some degree, justifiable (Martínez, Sánchez, and Galindo 2022) (McGee, Benk, and Yüzbaşı 2015).

To ascertain the most significant predictors we instantiated a classification decision tree which was resampled through k-fold cross-validation to obviate issues of overfitting (Prajwala 2015). Decision trees split data into progressively smaller categories through “binary classification judgments” (Belcic 2024) from which a variable importance plot can be created to visualise the main predictors of divorce attitudes. Decision trees are considered “powerful and popular tools for classification and prediction”, their strengths lying in their interpretability, robustness to the presence of outliers (Prajwala 2015), and classification reliability and accuracy (Podgorelec et al. 2002).

2.4.2 Investigating Predictors of Divorce - The Process

First a new dataframe was created containing all desired predictors, as well as the chosen outcome variable.

# need to make new dataframe with all desired predictors and outcome as binary factor variable

wvs_dt_df <- wvs_qs %>%
  select(-country_name, 
         -B_COUNTRY, 
         -B_COUNTRY_ALPHA, 
         -N_REGION_ISO, 
         -N_REGION_WVS, 
         -O1_LONGITUDE, 
         -O2_LATITUDE) %>%
  # Create binary outcome variable for Divorce 
  mutate(outcome_185 = factor(ifelse(Q185 == 1, 0, 1)))

This dataframe was subsequently split into training and test sets (Podgorelec et al. 2002), in which 75% was allocated to the former, and the outcome variable was stratified so each class will be proportionally represented in each set. Cross-validation folds were then specified (k = 10). K-fold cross-validation is a preferable means of determining the likelihood an “outcome is generated by chance” (Gorriz et al. 2024), allows for hyperparameter tuning (Anguita et al. 2012), and outperforms standard hypothesis testing (Gorriz et al. 2024) and leave-one-out cross-validation (Rao, Fung, and Rosales 2008). The number of folds (k = 10) was chosen based on the premise that larger k values risk underfitting and vice versa for smaller k values (Agrawal 2023). Further, researchers have posited standard k values as 5, 10, and 20 (Anguita et al. 2012), from which k = 10 was selected as it arguably mitigates the risk of over- and under-fitting. This is considered a typical k value as it reflects the “trade-off” between bias and variance in error estimation (Gorriz et al. 2024) and support for k = 10 when employing decision trees is cited in the literature (Rao, Fung, and Rosales 2008). Finally, the evaluation metric was defined, in this case we selected the area under the receiver operator characteristic curve, otherwise termed ROC AUC.

The ROC curve illustrates the interplay between the true and false positive rates whereby better classification models have higher true positive rates and curves on the left of the graph, the area below which quantifies the performance of the curve (Vujovic 2021). This area score measures the likelihood a “randomly selected positive has a higher confidence score” than a negative counterpart (Belcic 2024). Higher ROC AUC scores (closer to 1) indicate “excellent” model performance, scores around 0.5 approximate random estimation, and scores closer to 0 indicate “poor” model performance (Prasanna 2024) and denotes that the classification model scores negatives with greater probabilities than positives (Belcic 2024). Both the curve and the area are considered “important” classification metrics (Prasanna 2024), and are beneficial in determining model performance in terms of “ranking predictions” as well as when data is imbalanced (Czakon 2022), which distorts metrics such as accuracy (Prasanna 2024).

# Create training and test sets
set.seed(123) 
wvs_split <- initial_split(wvs_dt_df, prop = 0.75, strata = outcome_185)
wvs_train <- training(wvs_split) 
wvs_test <- testing(wvs_split)

# Create cross-validation folds
wvs_cv <- vfold_cv(wvs_train, v = 10)

# Set evaluation metrics
eval_metrics <- metric_set(roc_auc)

A decision tree model was created, specifying the tunable hyperparameters - cost complexity, tree depth, minimum node size - the model type, classification, and package used to fit the model, rpart. A recipe was written to define the outcome, divorce, and predictors. The included predictors were settlement size and type, urban-rural dimension, important to life values - family, friends, leisure time, politics, work, religion - sex, age, citizenship, whether respondents have children, education level, social class (subjective), and religious denomination. These predictors represent the aforementioned factors that impact attitude formation, as well as other types of regional definition (e.g., urban) which could enhance our understanding of the differences observed. We decided to omit the other actions (Q177-195), as we were trying to determine characteristics that predict regional differences in divorce justifiability and answers to these questions would not yield that result. A workflow combined the model and recipe.

# Create a cross-validated decision tree model
wvs_dt_model <-
  decision_tree(
    cost_complexity = tune(),
    tree_depth = tune(),
    min_n = tune()
  ) %>%
  set_mode("classification") %>%
  set_engine("rpart")

# Create recipe
wvs_dt_recipe <-
  wvs_train %>%
  recipe(outcome_185 ~ .) %>%
  step_rm(Q177, Q178, Q179, Q180, Q181, Q182, Q183, Q184, Q185, Q186, Q187, Q188, Q189, Q190, Q191, Q192, Q193, Q194, Q195)

# Create a workflow for cross-validated model
wvs_dt_workflow <-
  workflow() %>%
  add_recipe(wvs_dt_recipe) %>%
  add_model(wvs_dt_model)

We tuned the hyperparameters according to suitable standards for an average-sized dataset, cost complexity (0.001-0.1), tree depth (3-10), and minimum node size (1-20).

# Create a hyperparameter tuning grid
wvs_dt_grid <- expand_grid(
  cost_complexity = c(0.001, 0.01, 0.1),
  tree_depth = c(4, 6, 8),
  min_n = c(5, 10, 20)
)

# Grid search using parallel processing 
doParallel::registerDoParallel()

n_cores <- parallel::detectCores() - 1

cl <- makeCluster(n_cores)
registerDoParallel(cl)

wvs_dt_fit <-
  tune_grid(
    wvs_dt_workflow,
    resamples = wvs_cv,
    metrics = eval_metrics,
    grid = wvs_dt_grid
  )

stopCluster(cl)

Hyperparameters

We then plotted a tuning grid which revealed that according to the best ROC AUC value, the model where cost complexity is 0.001, tree depth is 8, and minimum node size is 5, was the optimal classification model.

# Plot hyperparameters
wvs_dt_fit %>% 
  collect_metrics() %>%
  filter(.metric == "roc_auc") %>%
  ggplot(aes(x = as.factor(cost_complexity), y = mean, group = tree_depth, color = as.factor(tree_depth))) +  geom_point() +
  geom_line(linetype = "dashed") +
  theme_minimal() +
  labs(x = "Cost Complexity", y = "Mean Cross-Validated ROC_AUC", color = "Tree Depth") +
  ggtitle("Cross Validated Decision Tree Hyperparameter Tuning")

# Select best model based on R^2
highest_roc_dt <- select_best(wvs_dt_fit, metric = "roc_auc")
highest_roc_dt %>% kable()

cost_complexity	tree_depth	min_n	.config
0.001	8	5	Preprocessor1_Model07

This model was fitted to the training set and evaluated on the test set. Unseen test data is used for model evaluation to protect against overfitting and inaccurate predictions (Belcic 2024).

# Fit best model to training set
wvs_final_fit <- 
  wvs_dt_workflow %>% 
  finalize_workflow(highest_roc_dt) %>% 
  fit(wvs_train)

2.4.3 Investigating Predictors of Divorce - The Outcome

The ROC AUC score of over 0.5 indicates that the classification predictions made exceed random estimation (0.5), yet are not perfect indicating a margin of error in model predictions.

# Evaluate best model on test set
wvs_final_fit %>%
  augment(new_data = wvs_test) %>%
  eval_metrics(truth = outcome_185, estimate = '.pred_0', '.pred_1', event_level = "second") %>% kable()

.metric	.estimator	.estimate
roc_auc	binary	0.6701818

The decision tree was then visualised however, due to the default formatting of the rpart.plot function it was not easily interpretable. Considering this, and that the objective of this analysis was to determine the most important predictors of divorce justifiability, we created a variable importance plot with the vip function.

This visualisation yielded that, as we expected, religiosity was by far the most important predictor of divorce justifiability. The following five predictors were similar in terms of importance: valuing work, urban-rural, settlement size, settlement type, and education level. Surprisingly social class and valuing of politics were marginally important predictors, converse to research findings of an association between political conservatism and economic and moral issues (Stack, Adamczyk, and Cao 2010) and national wealth predicting “evaluations of divorce” (Toth and Kemmelmeier 2009).

Findings would likely differ if the model was applied to different outcomes, such as euthanasia justifiability, which is notably affected by age, socioeconomic status, urban-rural differences (Tutić Grokša et al. 2025), as well as religious denominations (Karumathil & Tripathi, 2022). Despite this, we could arguably expect religiosity to remain an important predictor for many of the action justifiability values as it has been related to euthanasia (Inglehart et al. 2021), homosexuality (Furnham and Saito 2009), abortion (Tutić Grokša et al. 2025), and suicide (Lester 2003). The sexual behaviours - premarital and casual sex - are seemingly more influenced by sexual regulation (Broude 1975).

Statement on Use of Generative AI Tools, As Per School and Course-Specific Policies

In response to the guidance on the full authorised use of generative AI in assessment, we have utilised generative AI tools throughout the learning and assessment process as permitted. These tools were used in the following ways while making sure we used these tools responsibly to ensure our learning and development align with course expectations. Under Descriptive Statistics, we used AI to help us develop clean and insightful graphs, linking to the question_labels created to better display out outputs, and to investigate the alternate measure of the geometric mean Under Clustering Actions, we used AI to implement the process of investigating Jensen-Shannon divergence, and later in Part 2, used AI to aid us with parallel processing due to the large amount of data involved Furthermore, in this section we used AI to print the outputs of agglomerative hierarchical clustering for our desired k value Additionally, we used AI to determine the point of maximum curvature, and alternative points of curvature (seen in the graph which was generated by AI) to help us define an appropriate eps value. Under Part 2.1 Country Means, we used AI to convert our graphs displaying the geometric mean per question, per country and national-global mean divergence into an interactive plot Furthermore, geographic maps were largely aided by AI Under Part 2.4 Prediction, we used AI to create the binary classification of justification (where “Never Justifiable” became 0, and anything greater than that became 1) and to fix issues with evaluating the model on the test set using ROC AUC Within our models, we used AI to suggest suitable parameters and aid with hyperparameter tuning (e.g. tree_depth, minPts, eps, cost_complexity) To avoid any undue repercussions, we have stated throughout the sections of this assignment where we were aided by ChatGPT, to the best of our ability to ensure transparency and adherence to the guidelines while mitigating any concerns.

References

Agrawal, Saurav. 2023. “Hyperparameter Tuning of KNN Classifier. Medium.” June 4, 2023. https://medium.com/@agrawalsam1997/hyperparameter-tuning-of-knn-classifier-a32f31af25c7.

Almosaed, Nora. 2004. “Violence Against Women: A Cross‐cultural Perspective.” Journal of Muslim Minority Affairs 24 (1): 67–88. https://doi.org/10.1080/1360200042000212124.

Anguita, Davide, Luca Ghelardoni, Alessandro Ghio, Luca Oneto, and Sandro Ridella. 2012. “The ‘k’ in k-Fold Cross Validation.” Computational Intelligence.

Babb, Alexius Q. 2015. “Stealing Is Stealing: A Cross-Cultural Analysis on the Perception of Criminality for Physical Theft Versus Illegal Downloading.” IU Journal of Undergraduate Research 1 (1): 54–59. https://doi.org/10.14434/iujur.v1i1.13320.

Belcic, Ivan. 2024. “What Is Classification in Machine Learning? IBM.” October 15, 2024. https://www.ibm.com/think/topics/classification-machine-learning.

Broude, Gwen J. 1975. “Norms of Premarital Sexual Behavior: A Cross-Cultural Study.” Ethos 3 (3): 381–402. https://www.jstor.org/stable/640100.

Chon, Don Soo. 2015. “Gender Equality, Liberalism and Attitude Toward Prostitution: Variation in Cross-National Study.” Journal of Family Violence 30 (7): 827–38. https://doi.org/10.1007/s10896-015-9713-y.

Colucci, Erminia, and David Lester. 2012. Suicide and Culture: Understanding the Context. Hogrefe Publishing GmbH.

Czakon, Jakub. 2022. “F1 Score Vs ROC AUC Vs Accuracy Vs PR AUC: Which Evaluation Metric Should You Choose? Neptune.ai.” July 21, 2022. https://neptune.ai/blog/f1-score-accuracy-roc-auc-pr-auc.

Ember, Carol R., and Melvin Ember. 2005. “Explaining Corporal Punishment of Children: A Cross-Cultural Study.” American Anthropologist 107 (4): 609–19. https://doi.org/10.1525/aa.2005.107.4.609.

Furnham, Adrian, and Kaoru Saito. 2009. “A Cross-Cultural Study of Attitudes Toward and Beliefs about, Male Homosexuality.” Journal of Homosexuality 56 (3): 299–318. https://doi.org/10.1080/00918360902728525.

Gorriz, Juan M., R. Martin Clemente, F. Segovia, J. Ramirez, A. Ortiz, and J. Suckling. 2024. “Is k-Fold Cross Validation the Best Model Selection Method for Machine Learning?” arXiv. https://doi.org/10.48550/arXiv.2401.16407.

Green, Eva G. T., Jean-Claude Deschamps, and Dario Páez. 2005. “Variation of Individualism and Collectivism Within and Between 20 Countries: A Typological Analysis.” Journal of Cross-Cultural Psychology 36 (3): 321–39. https://doi.org/10.1177/0022022104273654.

Haerpfer, C., R. Inglehart, A. Moreno, C. Welzel, K. Kizilova, J. Diez-Medrano, M. Lagos, P. Norris, E. Ponarin, and B. Puranen, eds. 2022. “World Values Survey: Round Seven – Country-Pooled Datafile Version 6.0.” https://doi.org/https://doi.org/10.14281/18241.24.

Inglehart, Ronald C., Ryan Nash, Quais N. Hassan, and Judith Schwartzbaum. 2021. “Attitudes Toward Euthanasia: A Longitudinal Analysis of the Role of Economic, Cultural, and Health-Related Factors.” Journal of Pain and Symptom Management 62 (3): 559–69. https://doi.org/10.1016/j.jpainsymman.2021.01.009.

Karumathil, Anjana A., and Ritu Tripathi. 2022. “Culture and Attitudes Towards Euthanasia: An Integrative Review.” OMEGA - Journal of Death and Dying 86 (2): 688–720. https://doi.org/10.1177/0030222820984655.

Lee, Yueh-Ting, Russ Kleinbach, Pei-Cheng Hu, Zu-Zhi Peng, and Xiang-Yang Chen. 1996. “Cross-Cultural Research on Euthanasia and Abortion.” Journal of Social Issues 52 (2): 131–48. https://doi.org/10.1111/j.1540-4560.1996.tb01572.x.

Lee, Yueh-Ting, Seiji Takaku, Victor Ottati, and Gonggu Yan. 2004. “Perception and Interpretation of Terrorism, Justice, and Conflict: Three Cultures and Two Sides of One Coin.”

Lester, David. 2003. “Cross-Cultural Perspectives on Suicide.” In Handbook of Death and Dying. SAGE.

Loll, Dana, and Kelli Stidham Hall. 2019. “Differences in Abortion Attitudes by Policy Context and Between Men and Women in the World Values Survey.” Women & Health 59 (5): 465–80. https://doi.org/10.1080/03630242.2018.1508539.

Martínez, Pilar Rodríguez, Antonio Jesús Segura Sánchez, and Carmelo Jesús Aguilera Galindo. 2022. “Justification of Terrorism According to World Values Survey (2017–2020).” Research in Globalization 5 (December): 100085. https://doi.org/10.1016/j.resglo.2022.100085.

McCarthy, Daniel, and Ian Brunton-Smith. 2024. “Attitudes Towards the Death Penalty: An Assessment of Individual and Country-Level Differences.” European Journal of Criminology 21 (1): 116–39. https://doi.org/10.1177/14773708221097670.

McGee, Robert W., Serkan Benk, and Bahadır Yüzbaşı. 2015. “Religion and Ethical Attitudes Toward Accepting a Bribe: A Comparative Study.” Religions 6 (4): 1168–81. https://doi.org/10.3390/rel6041168.

Navot, Doron, Yaniv Reingewertz, and Nissim Cohen. 2016. “Speed or Greed? High Wages and Corruption Among Public Servants.” Administration & Society 48 (5): 580–601. https://doi.org/10.1177/0095399715623313.

Podgorelec, Vili, Peter Kokol, Bruno Stiglic, and Ivan Rozman. 2002. “Decision Trees: An Overview and Their Use in Medicine.” Journal of Medical Systems 26 (5): 445–63. https://doi.org/10.1023/A:1016409317640.

Prajwala, T. R. 2015. “A Comparative Study on Decision Tree and Random Forest Using r Tool.” IJARCCE, January, 196–99. https://doi.org/10.17148/IJARCCE.2015.4142.

Prasanna, Chanaka. 2024. “Understanding ROC and AUC in Model Evaluation. Medium.” July 19, 2024. https://medium.com/@chanakapinfo/understanding-roc-and-auc-in-model-evaluation-88fe4ac7dfc7.

Rao, R. Bharat, Glenn Fung, and Romer Rosales. 2008. “On the Dangers of Cross-Validation. An Experimental Evaluation.” In Proceedings of the 2008 SIAM International Conference on Data Mining (SDM), 588–96. Proceedings. Society for Industrial; Applied Mathematics. https://doi.org/10.1137/1.9781611972788.54.

Roberts, Louisa L. 2019. “Changing Worldwide Attitudes Toward Homosexuality: The Influence of Global and Region-Specific Cultures, 1981–2012.” Social Science Research 80 (May): 114–31. https://doi.org/10.1016/j.ssresearch.2018.12.003.

Stack, Steven, Amy Adamczyk, and Liqun Cao. 2010. “Survivalism and Public Opinion on Criminality: A Cross-National Analysis of Prostitution.” Social Forces 88 (4): 1703–26. https://doi.org/10.1353/sof.2010.0029.

Tausch, Arno. 2019. “Multivariate Analyses of the Global Acceptability Rates of Male Intimate Partner Violence (IPV) Against Women Based on World Values Survey Data.” The International Journal of Health Planning and Management 34 (4): 1155–94. https://doi.org/10.1002/hpm.2781.

Torgler, Benno, and Friedrich Schneider. 2005. “Attitudes Towards Paying Taxes in Austria: An Empirical Analysis.” Empirica 32 (2): 231–50. https://doi.org/10.1007/s10663-004-8328-y.

Torgler, Benno, and Neven T. Valev. 2010. “Gender and Public Attitudes Toward Corruption and Tax Evasion.” Contemporary Economic Policy 28 (4): 554–68. https://doi.org/10.1111/j.1465-7287.2009.00188.x.

Toth, Katalin, and Markus Kemmelmeier. 2009. “Divorce Attitudes Around the World: Distinguishing the Impact of Culture on Evaluations and Attitude Structure.” Cross-Cultural Research 43 (3): 280–97. https://doi.org/10.1177/1069397109336648.

Tutić Grokša, Ivana, Ana Depope, Tijana Trako Poljak, Igor Eterović, Toni Buterin, Robert Doričić, Mariana Gensabella, et al. 2025. “Students’ Attitudes Toward Euthanasia and Abortion: A Cross-Cultural Study in Three Mediterranean Countries.” BMC Medical Ethics 26 (1): 9. https://doi.org/10.1186/s12910-025-01167-8.

Vujovic, Željko Ð. 2021. “Classification Model Evaluation Metrics.” International Journal of Advanced Computer Science and Applications (IJACSA) 12 (6). https://doi.org/10.14569/IJACSA.2021.0120670.

World Values Survey. n.d. “WVS Database. World Values Survey.” Accessed February 16, 2025. https://www.worldvaluessurvey.org/WVSContents.jsp.

Wormald, Benjamin. 2014. “Religion in Latin America. Pew Research Center.” November 13, 2014. https://www.pewresearch.org/religion/2014/11/13/religion-in-latin-america/.

Zhang, Zhiyuan, Guoxin Ni, and Yanguo Xu. 2020. “Comparison of Trajectory Clustering Methods Based on k-Means and DBSCAN.” In 2020 IEEE International Conference on Information Technology,big Data and Artificial Intelligence (ICIBA), 1:557–61. https://doi.org/10.1109/ICIBA50161.2020.9277214.